Single-Provider Enterprises Fell From 64% to 28% in a Quarter. The Buyer Picks the Model Now. Ship the Label Set.

Databricks: one-provider enterprises went from 64% to 28% in a quarter and the top-scoring model fell to 0.6% of tokens. 150 labels built my router in 3 days.

Falk Gottlob · · 6 min read

Foundation Falkster cover on pink: a navy hand flips a cream three-position switch, and beside it sits a stack of cream index cards with an orange tag tied on by string.

The Databricks AI Index landed with a headline about open-weight models. The number that matters more for anyone selling an AI product is about who is doing the switching. These are Field Notes on that number.

The short version

  • Databricks published its AI Index on November 16, 2026, from token traffic across more than 10,000 companies on its gateway.
  • Enterprises relying on a single model provider fell from about 64% in July to 28% by November 2. Open-weight share went from 7.3% to 30.9% of enterprise tokens.
  • Claude Opus 5.5 took 37.9% of proprietary tokens in six weeks, and the top-scoring model on Databricks' benchmark fell to 0.6%. Jason Lemkin's read on SaaStr is that "built on" one model is no longer a moat and that founders should build the routing layer.
  • The index says the routing is already happening on the buyer's side. So the vendor's durable asset is a labeled set of its own requests showing which can move to a cheaper model and which cannot.
  • At Smartcat, 150 labeled translation requests gave engineering a working router in three days.

The numbers, and who reported them

I read Lemkin's post in full and checked the figures against a fetched summary of Databricks' own post. Lemkin's caveat comes first: Databricks hosts open-weight models and has a commercial interest in companies using them, so its customers probably lean further that way than the average enterprise.

From the index, as both report it:

  • One provider. About 64% of enterprises on the gateway relied on a single model provider in July. By November 2 it was 28%.
  • Open-weight. 7.3% of enterprise token traffic in July, 30.9% in November.
  • A new model, fast. Claude Opus 5.5 went from 7.2% to 37.9% of proprietary-model tokens in six weeks. Anthropic's overall share rose from 57.6% to 68.8% over that stretch.
  • The best model, shrinking. Claude Fable 5.1 tops Databricks' coding benchmark at 96.0% task success and $7.47 per task. Its share of proprietary tokens went from 5.5% to 0.6%.
  • Cost per task. GPT-6 Luna and Claude Opus 5.5 both scored 89.1%, at $0.05 and $1.62 per task. Gemini 3.8 Flash averaged 140 turns and $11.97 per task on a low token price.
  • Spend per user. A median of $2.49 and a 99th percentile of $699.74, about 281x. Lemkin reads these as weekly figures.

Lemkin draws founder advice from each. Rebuild cost of goods on cheaper models for bulk work. Stop saying "we're built on" a single lab. Measure cost per completed task. Stop pricing flat per seat. Build the routing layer now.

Who is routing

That last piece of advice is where I'd turn the page around.

Every number above is a buyer moving. An enterprise went from one provider to several. An enterprise sent bulk work to open-weight. An enterprise moved a third of its proprietary traffic to a new model in six weeks, which Lemkin attributes to a gateway turning the switch into a config change. Databricks, he notes, did the same thing internally and stayed on an older Opus because the upgrade wasn't worth the cost for its workloads.

So the routing layer exists. It belongs to the customer, and it sits in front of whatever you sell them. Lemkin says as much in one line: buyers want to know you can use whichever model they've already approved, including the open-weight ones running inside their own perimeter.

For a product team that changes what "our model strategy" means. The model on your architecture slide is a default. A customer's platform team can change it on a Tuesday without telling you, and your product's quality on Wednesday is yours to explain.

What the vendor still owns

If the buyer holds the switch, the vendor's asset is knowing what the switch does to its own product. Request type by request type: this one is fine on the cheapest tier, this one needs the middle, and this one breaks below the top.

That knowledge is a labeled set. In Label Before You Route I described the one I built at Smartcat. The decision was whether a translation request should stay with AI, go to a human linguist, or be escalated. I pulled 150 real requests from the previous quarter and labeled each with where it should have gone. Engineering took the set and had a working router in three days. We ran it daily, and every regression pointed at the exact inputs that failed.

The index's benchmark table is the argument for doing this on your own requests. Two models tie at 89.1% and differ about 32x in cost per task. One cheap-per-token model takes 140 turns. A public benchmark built from Databricks' pull requests cannot tell you which of those results holds for your tickets or your contracts.

Three things I'd have ready before a customer asks:

  1. The set. About 150 real requests, labeled with the lowest tier you'd trust for each.
  2. The floor, in writing. Which request types you will not serve below a named tier, and what the customer sees if their approved models are all below it.
  3. The rerun. The set runs on every model the customer approves, and the result goes to them.

The second one is also a margin document. In Anthropic's S-1 Has No Gross Margin Line. Your AI Product Needs One. I argued that every cost-per-task cut needs a decision about who keeps it. When the customer chooses the cheaper model, they've made that decision for you unless your price is tied to the outcome and not the model.

This week

Pull 150 real requests from last quarter and label each with the cheapest tier you'd trust. Then run them on one model you don't use today that a customer has asked about. This sits in SaaS to AI Business Models because a product whose model is the customer's setting has to price and promise on something else.

Related answer: What should an AI product vendor do when enterprise buyers choose the model?

Sources: Jason Lemkin, "5 Interesting Learnings from the Databricks AI Index," SaaStr, November 18, 2026, read in full from the feed: the five learnings, the caveat, and the founder advice. Databricks, "Introducing the Databricks AI Index: The state of enterprise AI tokenomics," November 16, 2026, read as a fetched summary that confirmed the 64% and 28%, 7.3% and 30.9%, 57.6% and 68.8%, 7.2% and 37.9%, 5.5% and 0.6%, the benchmark scores and costs per task, the $2.49 and $699.74, and the 10,000 companies. Figures in Lemkin's post that the summary did not confirm are not used. The 150 requests and three days are mine, from the post linked above.

64% to 28%

of enterprises on Databricks' gateway relying on a single model provider, July to November 2, 2026, per the Databricks AI Index.
Share this post

Frequently asked

What is the Databricks AI Index?+

A report Databricks published on November 16, 2026, based on token traffic from more than 10,000 companies that route AI through its Unity Gateway. As summarized from the Databricks post, the data covers September 21 to November 7, 2026. Jason Lemkin wrote up five learnings from it on SaaStr on November 18 and notes that Databricks hosts open-weight models and has a commercial interest in their use.

Which numbers moved?+

Per the index: enterprises relying on one model provider fell from about 64% in July to 28% by November 2. Open-weight share of enterprise tokens rose from 7.3% to 30.9% between July and November. Claude Opus 5.5 went from 7.2% to 37.9% of proprietary-model tokens in six weeks, while Claude Fable 5.1, the top scorer on Databricks' coding benchmark, fell from 5.5% to 0.6%.

Why does this mean the buyer picks the model?+

Because the switching is happening on the enterprise side, through a gateway, at a speed vendors do not control. Lemkin's reading is that buyers want to know a product can use whichever model they have already approved, including open-weight ones inside their own perimeter. That makes the model a setting the customer holds.

What is the label set?+

A set of real requests from your product, each labeled with the cheapest model tier you would trust with it, or with which requests cannot move at all. At Smartcat I labeled 150 real translation requests by where each should have gone, and engineering had a working router in three days. The method is in Label Before You Route.

What does cost per task have to do with it?+

It is why the label has to be per request type. On Databricks' benchmark, GPT-6 Luna and Claude Opus 5.5 both scored 89.1%, at $0.05 and $1.62 per task. Gemini 3.8 Flash had a low token price and averaged 140 turns and $11.97 per task. A tie on a benchmark says nothing about your requests.

Did you read the index itself?+

I read Lemkin's SaaStr post in full and a fetched summary of the Databricks blog post, which confirmed the figures used here. The figures Lemkin cites that the summary did not confirm, including the harness comparison and average models per customer, are left out.

DatabricksDatabricks AI IndexJason LemkinSaaStrmodel routingopen-weight modelsgross margincost per task
Helpful?

Keep reading

The short answer

Engineering Builds the Substrate, Not FeaturesIn an AI-native org, engineering's highest-leverage work is the substrate that lets PMs and designers ship safely: guardrails, evals, isolated deploys.Read the chapter

Part of

Enterprise AI Agents

About the author

Falk Gottlob

Falk Gottlob

Product Executive · Founder, Falkster.AI

Thirty years shipping product, from Microsoft Research and Adobe to Salesforce, where he grew Quip into what became Slack Canvas. Four startup exits, including a $6.5B healthcare merger and a company Microsoft bought. Four-time Chief Product Officer. Now founder of Falkster.AI, building Heidi, the company brain for AI-native businesses. This notebook is written from inside the build, not above it.

Sign in with LinkedIn to leave a comment.

Sign in with LinkedIn